UAV-Assisted TinyML Transfer Learning for Edge Soil-Moisture Forecasting in Urban and Peri-Urban Smart Farming
Journal of Urban Development and Smart Cities · 01.12.2024 · 1 · 73-83
DOI: 10.66033/judsc2024-108 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Urban Development and Smart Cities, online erschienen am 01.12.2024. Verfasst von Wenping Wang und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Distributed irrigation intelligence is increasingly relevant to smart-city systems because urban and peri-urban food production now depends on reliable sensing, low-power communications, and resource-efficient water management. This paper presents a UAV-assisted tiny machine learning (TinyML) framework for edge soil-moisture forecasting using transfer learning on low-power internet of things nodes. The system combines bespoke ESP32-based sensing hardware, unmanned aerial vehicle (UAV)-enabled over-the-air model delivery, and lightweight deep learning inference for local decision support. The reported implementation uses two datasets: a five-site public dataset collected at 10-minute intervals over three years, and a three-month field dataset collected in Amman, Jordan, using a custom capacitive soil-moisture platform. The forecasting pipeline applies hourly aggregation, interpolation of sparse missing values, min-max scaling, and seasonal-trend decomposition using Loess before training compact deep neural network (DNN) and long short-term memory (LSTM) models. The transfer-learning experiment shows that only 441 trainable parameters are updated out of 5,293 total DNN parameters; without transfer learning the model needs more than 300 epochs to converge, whereas transfer learning reduces convergence to fewer than 25 epochs on average and achieves an R2 of 92.9%. In the reported edge deployment case, a compressed DNN with architecture 40 × 20 × 10 × 1 occupies 7,920 bytes, produces inference in 97.80 ms, attains an average R2 of 97.13% with an MSE of 0.0036, and can be transferred by over-the-air update in under 10 s. An 8-unit LSTM reaches 99.8% average R2 with an MSE of 7.9228 × 10−5, while larger LSTM configurations in the full performance sweep deliver validation R2 values above 99.9%. Framed for smart-city and urban development scholarship, the study demonstrates that edge-native irrigation intelligence can reduce communication burden, improve resilience in connectivity-constrained environments, and support data-driven water stewardship in distributed urban agriculture.
Zitieren
APA
Wang, W., & Brühl, R. (2024). UAV-Assisted TinyML Transfer Learning for Edge Soil-Moisture Forecasting in Urban and Peri-Urban Smart Farming. Journal of Urban Development and Smart Cities, 1, 73-83. https://doi.org/10.66033/judsc2024-108
Vancouver
Wang W, Brühl R. UAV-Assisted TinyML Transfer Learning for Edge Soil-Moisture Forecasting in Urban and Peri-Urban Smart Farming. Journal of Urban Development and Smart Cities. 2024;1:73-83. doi:10.66033/judsc2024-108
BibTeX
@article{Wang2024,
author = {Wang, Wenping and Brühl, Rolf},
title = {UAV-Assisted TinyML Transfer Learning for Edge Soil-Moisture Forecasting in Urban and Peri-Urban Smart Farming},
journal = {Journal of Urban Development and Smart Cities},
year = {2024},
volume = {1},
pages = {73--83},
doi = {10.66033/judsc2024-108}
}
EUVIMED